Learning-Based Object Identification and Segmentation Using Dual-Energy CT Images for Security.

Area of Science:

• Imaging Science
• Computer Vision
• Security Technology

Background:

• Dual-energy X-ray computed tomography (DECT) is crucial for nondestructive evaluation in baggage screening.
• Current methods struggle with object segmentation and identification due to image artifacts, clutter, and metal interference.
• Existing approaches often decouple segmentation and material identification, limiting accuracy.

Purpose of the Study:

• To develop a novel, learning-based framework for joint object segmentation and identification from DECT data.
• To enhance the robustness of threat object detection against metal artifacts, image noise, and clutter.
• To improve the accuracy and reliability of automated baggage screening systems.

Main Methods:

• A learning-based framework for joint segmentation and identification directly from volumetric DECT images.
• Incorporation of data weighting to mitigate metal artifacts and an object boundary field to reduce splitting.
• Formulation as a multilabel discrete optimization problem solved via an efficient graph-cut algorithm.

Main Results:

• The proposed method demonstrates robustness to severe image noise, artifacts, and clutter.
• Accurate segmentation and identification of objects of interest were achieved without splitting.
• Successful testing on real DECT data confirmed the framework's potential for practical application.

Conclusions:

• The novel framework offers a significant advancement in automated threat object detection using DECT.
• Joint segmentation and identification effectively address limitations of conventional decoupled approaches.
• This technology has the potential to improve airport security screening efficiency and accuracy.